Spike-Guided Multi-Stage RUL Estimation via Physics-Constrained Temporal Networks

Predicting the remaining useful life (RUL) of machines is important for safe and efficient operation, but short signal spikes near the start of faults often disturb prediction results. This study proposes a spike-guided, multi-stage method that combines data learning with a simple physical update. A convolution block filters spike signals using envelope energy and basic variance checks, and a two-path predictor joins time features with a physical correction term. Tests on the NASA turbofan and PRONOSTIA bearing datasets showed that the method cut RMSE by 14–21% and raised the early-warning score by about 18% compared with other deep learning models. The spike check step also reduced false alarms and kept the trend of wear smoother over time. These results show that short bursts hold key signs of early faults and that adding physical rules helps keep forecasts more stable. The method can help with early maintenance planning in factory systems, though wider tests under more working conditions are still needed.